Flink Pure
Streaming
Paco Guerrero
Big Data & Solutions Architect 9/21/16
Not for
Geeks
Life as Time
4
Anything as Time
Flink as Time
Streaming vs Batch
7
“Abstraction of reality used to facilitate information processing”
Batch
Batch
Batch
Batch
All
Input
Batch
Batch Job
All
Input
Batch
Batch Job
All
Input
All
Output
Nothing about time
Timestamps used as trick to
keep real time fingerprint
Streaming
“Continuous processing of data that is continuously produced”
Streaming
“Streaming is the next programming paradigm for
data applications, and you need to start thinking in
terms of streams”
“Continuous processing of data that is continuously produced”
Streaming
“Streaming is the next programming paradigm for
data applications, and you need to start thinking in
terms of streams”
“Continuous processing of data that is continuously produced”
Data Stream: Infinite sequence of data arriving in a
continuous fashion.
Streaming
“Streaming is the next programming paradigm for
data applications, and you need to start thinking in
terms of streams”
“Continuous processing of data that is continuously produced”
Data Stream: Infinite sequence of data arriving in a
continuous fashion.
Stream processing is the backbone of the new data
infrastructure.
Streaming
“Streaming is the next programming paradigm for
data applications, and you need to start thinking in
terms of streams”
“Continuous processing of data that is continuously produced”
Data Stream: Infinite sequence of data arriving in a
continuous fashion.
Stream processing is the backbone of the new data
infrastructure.
“The world beyond batch”
A high-level tour of modern data-processing concepts. By Tyler Akidau
August 5, 2015 https://www.oreilly.com/ideas/the-world-beyond-batch-streaming-101
Streaming
Streaming
Streaming Job
Streaming
Streaming Job
Streaming
Streaming Job
Real Life Time !!
Streaming is the biggest change in
data infraestructure since Hadoop
Streaming
The biggest change is moving from
batch to streaming is handling time explicitly
Streaming
Micro Batch
Micro Batch
Micro Batch
Batch Job 1
Batch Job
Batch Job 2
All
Output
Batch Job 1
Micro Batch
Batch Job
Batch Job 2
All
Input
All
OutputBatch Job 3
All
Output
All
Output
Batch Job 1
Batch Frequency ?
Timestamps keeps real time
fingerprint
Micro Batch
Streaming Technologies
Batch StreamingMicro Batch
StateLess –
Record acknowledgements
CPU bounded performance
Not expressive declarative
functional API – Low Level API
Not auto scaling
Low level programmatic topology
Poor Streaming Windows
funcionalities
Not compatible with Hadoop APIs
Streams
Streaming Technologies
Batch StreamingMicro Batch
StateLess –
Record acknowledgements
CPU bounded performance
Not expressive declarative
functional API – Low Level API
Not auto scaling
Low level programmatic topology
Poor Streaming Windows
funcionalities
Not compatible with Hadoop APIs
Streams
Streaming Technologies
Batch StreamingMicro Batch
StateLess –
Record acknowledgements
CPU bounded performance
Not expressive declarative
functional API – Low Level API
Not auto scaling
Low level programmatic topology
Poor Streaming Windows
funcionalities
Not compatible with Hadoop APIs
Streams
Streaming Technologies
Batch StreamingMicro Batch
StateLess –
Record acknowledgements
CPU bounded performance
Not expressive declarative
functional API – Low Level API
Not auto scaling
Low level programmatic topology
Poor Streaming Windows
funcionalities
Not compatible with Hadoop APIs
Streams
Streaming Technologies
Streaming Technologies
Batch StreamingMicro Batch
Streaming Technologies
Batch StreamingMicro Batch
Streaming Technologies
Batch StreamingMicro Batch
Apache Flink
38
Flink
Open Source Stream Processing Framework.
Last available Release 1.1.1
Top Level Apache Project since Dec '14
Flink
Open Source Stream Processing Framework.
Last available Release 1.1.1
Top Level Apache Project since Dec '14
Main Features
Native Stream
Low Latency
High throughput
Stateful
Exactly-one guarantees
Distributed
Expressive APIs
And more ….
Flink
Open Source Stream Processing Framework.
Last available Release 1.1.1
Top Level Apache Project since Dec '14
Main Features
Native Stream
Low Latency
High throughput
Stateful
Exactly-one guarantees
Distributed
Expressive APIs
And more ….
Flink
Flink
Flink Integration
YARN upcoming..
.
Flink Integration
Flink Integration
YARN upcoming..
.
Flink Integration
YARN upcoming..
.
upcoming..
.
Flink Integration
YARN upcoming..
.
Flink Stack
Flink Runtime Engine
Flink Runtime
Engine
Distributed pipelined processing
Execute everything as Stream
Iterative ( cyclic ) dataflows
Mutable state in operations
Operate on managed memory (*)
Also works on batch !!
Job Manager
Client
Optimizer
Dataflow
Graph
Flink Runtime Engine
Distributed pipelined processing
Execute everything as Stream
Iterative ( cyclic ) dataflows
Mutable state in operations
Operate on managed memory (*)
Also works on batch !!
Workers ( Task Managers )
Job Manager
Client
Optimizer
Dataflow
Graph
Execution
Graph
Flink Runtime Engine
Stream Job
Batch Job
ML Job
Flink Runtime Engine
Graph Job
optimizer
optimizer
optimizer
optimizer
Stream Job
Batch Job
ML Job
Flink Runtime Engine
Graph Job
optimizer
optimizer
optimizer
optimizer
Tasks scheduled and executed in workers ( slots )
Tasks as chain of operators
Run operator logic in a pipelined fashion
Stream Job
Batch Job
ML Job
Flink Runtime Engine
Graph Job
optimizer
optimizer
optimizer
optimizer
Tasks scheduled and executed in workers ( slots )
Tasks as chain of operators
Run operator logic in a pipelined fashion
Stream Job
Batch Job
ML Job
Flink Runtime Engine
Graph Job
optimizer
optimizer
optimizer
optimizer
Tasks scheduled and executed in workers ( slots )
Tasks as chain of operators
Run operator logic in a pipelined fashion
Stream Job
Batch Job
ML Job
Flink Runtime Engine
Graph Job
optimizer
optimizer
optimizer
optimizer
If you want to know one thing about Flink
is that you don't need to know
the internals of Flink
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
lTime references
lOut of order events
lPowerful Windowing
Event Times & Windowing
Event Times & Windowing
Event
Time
Event
Time
Event Times & Windowing
Flink
Data Source
Event
Time
Event
Time
Ingestion
Time
Event Times & Windowing
Flink
Data Source
Flink
Window Operator
Event
Time
Event
Time
Ingestion
Time
Processing
Time
Event Time: when data is generated
Ingestion time: when data is loaded from source
Processing time: when data is processed
Event time help to process out- of-order events and replay elements as the ocurred (
deterministic results )
Explicit handling of time. 3
choices:
Event Times & Windowing
Event Times & Windowing
Event time. Out or Order
1 2 3 5 7
4 6 8 9 10
Event time. Out or Order
1 2 3 5 7
4 6 8 9 10
Event time. Out or Order
Out or Order
1 2 3 5 74 6 8 9 10
1 2 3 5 7
4 6 8 9 10 1 2 3 5 74 6 8 9 104
Event time. Out or Order
Ingestion Time WindowsOut or Order
1 2 3 5 74 6 8 9 10
1 2 3 5 7
4 6 8 9 10 1 2 3 5 74 6 8 9 10
1 2 3
4
4 5
Event time. Out or Order
6 7 8 9 10
Event Time Windows
Ingestion Time WindowsOut or Order
1 2 3 5 74 6 8 9 10
Event time. Watermarks
1 2 3 5 7
4 6 8 9 10
Event time. Watermarks
1 2 3 5 7
4 6 8 9 10
1 2 3 54 6 8
1 2 3 54 6 8
1 2 3
4
4 5
Event time. Watermarks
6 8
Event Time Windows
Ingestion Time WindowsOut or Order
1 2 3 5 7
4 6 8 9 10
1 2 3 54 6 8 910
1 2 3 54 6 8 910
1 2 3
4
4 5
Event time. Watermarks
6 8 9 10
Event Time Windows
Ingestion Time WindowsOut or Order
Not event time
before 5 will come
Late Time of 2
5
1 2 3 5 7
4 6 8 9 10
1 2 3 5 74 6 8 910
1 2 3 5 74 6 8 910
1 2 3
4
4 5
Event time. Watermarks
6 7 8 9 10
Event Time Windows
Ingestion Time WindowsOut or Order
Not event time
before 10 will
come
Late Time of 2
10
Windowing
Windows: grouping of events according to time, session*, count
Windowing
Windows: grouping of events according to time, session*, count
Powerful built-in windows:
Windowing
Windows: grouping of events according to time, session*, count
Powerful built-in windows:
Count: number of events to trigger the window. Process X last events each Y events.
Windowing
Windows: grouping of events according to time, session*, count
Powerful built-in windows:
Count: number of events to trigger the window. Process X last events each Y events.
Time:
lTumbling: trigger every X time with received events
lSliding: trigger every X time with received events in last Y time
Windowing
Windows: grouping of events according to time, session*, count
Powerful built-in windows:
Count: number of events to trigger the window. Process X last events each Y events.
Time:
lTumbling: trigger every X time with received events
lSliding: trigger every X time with received events in last Y time
Session: all events from session/user X until session time expired ( Gap )
Windowing
Windows: grouping of events according to time, session*, count
Powerful built-in windows:
Count: number of events to trigger the window. Process X last events each Y events.
Time:
lTumbling: trigger every X time with received events
lSliding: trigger every X time with received events in last Y time
Session: all events from session/user X until session time expired ( Gap )
High level API for user windows: Window Assigner, Trigger, Evictor
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
lManaged operator state for
backup/recovery
lSavepoints
Stateful Streaming
Op
Stateless Stream
Processing
Stateful Streaming
Op Op
State
Stateless Stream
Processing
Stateful Stream
Processing
lBuilt-in internal state in each operator for
exactly-once semantics
lUser state can be declared in each operator to be
saved locally in memory ( API, key/value pars )
lSnapshots: periodically local states
in memory are persisted in lightweight
distributed snapshots. No global pause !!
lCheckpoint as global consistent point-in-time
snapshot build by set of distributed snapshots.
lPluggable state backend for snapshots:
JobManager, HDFS, RocksDB
lSavepoints: user-triggered retained checkpoint
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
lExactly-once semantics with
managed operator state
lDistributed Snapshotting
Algorithm
Periodically
Chandy-Lamport Snapshots
“The global-state-detection algorithm is to be superimposed
on the underlying computation:
It must run concurrently with, but no alter, this underlying
computation”
. Triggers snapshots asynchronously
. Embedded snapshots algorithm in stream of data ( barriers )
. No global pause, lightweight impact in performance
Handling Checkpoints
Periodically
Chandy-Lamport Snapshots
“The global-state-detection algorithm is to be superimposed
on the underlying computation:
It must run concurrently with, but no alter, this underlying
computation”
. Triggers snapshots asynchronously
. Embedded snapshots algorithm in stream of data ( barriers )
. No global pause, lightweight impact in performance
Handling Checkpoints
Periodically
Chandy-Lamport Snapshots
“The global-state-detection algorithm is to be superimposed
on the underlying computation:
It must run concurrently with, but no alter, this underlying
computation”
. Triggers snapshots asynchronously
. Embedded snapshots algorithm in stream of data ( barriers )
. No global pause, lightweight impact in performance
Handling Checkpoints
snapshot
Job Manager
Periodically
pushes
barriers
for new state
New state X+1
Ack for Snapshot state X from Task N
Handling Checkpoints
snapshot
Job Manager
Handling Checkpoints
snapshot
Job Manager
Handling Checkpoints
snapshot
Job Manager
Handling Checkpoints
All Acks received
Register Checkpoint for restore
in case of fail
Streaming Fault Tolerance
In case of fail, last global checkpoint is recovered
( recovery from partial checkpoint / individual snapshots is coming )
Need of stateful source like kafka to ensure end-to-end exactly-once
semantic in case of fail.
Kafka sink doesn't guarantee end-to-end exactly-once ( multiple writes in
topic ) ( at least-once )
Semantics in Flink:
At Least Once: never loses events, events might be reprocessed
Exactly once: neither reprocessed nor lost events.
Exactly once by default, with low impact in performance
If you want to know one thing about Flink
is that you don't need to know
the internals of Flink
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
lPipelined runtime
lLatency vs throughput tunning
Exactly-once semantic with low impact in performance
Controllable checkpointing overhead
Higher throughput using processing time
Performance improvements thanks to:
. operator chaining during optimization phase
. own optimized serialization stack with code generation
Performance
Tunning
Benchmark for “Streaming Computation” published by Yahoo. Dec 18, 2015
https://yahooeng.tumblr.com/post/135321837876/benchmarking-streaming-computation-engines-at
Production use-case
lcounting ad impressions group by
campaign
laggregations over a 10 second
window
lsave current aggregate value to Redis
every second
Streaming
Benchmark
Throughput vs Latency Graph
Throughput ( 1000 events / sec )
99 Percentile
Latency ( ms )
Not Operator combinig in Storm, more complicate topology, more steps for events and more overhead
Apache Storm Without Trident
lAt least once / Double counting after fail / Lost state after Failures
lCPU bounded
Apache Spark
lLatency increase with throughput
Apache Flink
lExactly once / No double counting / No state loss
lLimited by bandwidth between Kafka and Flink cluster
l(1 GigE).
lkafka brokers within Kafka Cluster ( 10 GigE )
lAchieved 15 million messages /sec
l( before 3 million m/sec) with exactly once semantic
10,000,000 20,000,000
1 GigE
10 GigE
Performance
Tunning
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
lHigh Level API
lWide range of basic and advanced
operators
lJava , Scala. Python soon !!
API
API
Working on data streams ( bounded ? )
API
Working on data streams ( bounded ? )
Stream Processing: Explicit Handling of Time
API
Working on data streams ( bounded ? )
Stream Processing: Explicit Handling of Time
Java & Scala. Python coming.
Java: Bean type classes vs Tuples with position addresses.
Scala: case classes.
API
Working on data streams ( bounded ? )
Stream Processing: Explicit Handling of Time
Java & Scala. Python coming.
Java: Bean type classes vs Tuples with position addresses.
Scala: case classes.
Operators:
Sources: kafka, FileSystem, Cassandra …
Sinks: Kafka, HDFS, Cassandra ….
Transformations:
Basic: map, flatmap, filter, grouping, iterate, project, join, cross, …
Streaming: Windowing + Aggregations, Temporal Binary
Iterative Stream operators
API
Working on data streams ( bounded ? )
Stream Processing: Explicit Handling of Time
Java & Scala. Python coming.
Java: Bean type classes vs Tuples with position addresses.
Scala: case classes.
Operators:
Sources: kafka, FileSystem, Cassandra …
Sinks: Kafka, HDFS, Cassandra ….
Transformations:
Basic: map, flatmap, filter, grouping, iterate, project, join, cross, …
Streaming: Windowing + Aggregations, Temporal Binary
Iterative Stream operators
DataStream<?> DataSet<?>
Core API
1 implementation*, 2 interfaces
Source Map Reduce
Fliter
Join Sum Sink
Map
Source
Operators
Source Map Reduce
Fliter
Join Sum Sink
Map
Source
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Filter
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Filter
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Reduce
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Reduce
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Join
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Join
Operators
Source Map Reduce
Fliter
Join Sum Sink
Source
Operators
Events Time
&
Windows
Fault Tolerance
&
Correctness
State Handling
Low Latency
&
High Throughput
API Libraries SQL
Building Blocks
lEasy to use. SQL !!
lBased on Apache Calcite
API extension for DataSets y DataStreams
Based on relational Table abstraction
Table <=> Source / DataSet / DataStream
Operators like: where, select, as, groupBy, join, union, minus, distinct, orderBy, ...
Table API
Execute SQL-Like sentences on DataSets and Datastreams
Resuts returned as Table ( Table API ), convertible to DataStream or DataSets
SQL and Table API can be seamlessly mixed over DataStream/DataSets
Flink’s SQL support is not feature complete, yet.
Queries that include unsupported SQL will fail !!
SQL Support
SQL
Parsing and Logical plan for Table operators and SQL are optimized using Apache Calcite
Only supported a Subset of the comprehensive SQL standard
Apache Calcite provides with:
SQL Parsing
API for building expressions in relational algebra
Query planning engine
Provides SQL for Streaming Queries with windows aggregations
SELECT STREAM TUMBLE_END(rowtime, INTERVAL '1' HOUR) AS rowtime, productId, COUNT(*) AS c, SUM(units) AS units
FROM Orders
Apache Calcite
SQL Sentence
Apache Calcite:
SQL to Logical
Plan as Relational Algebra
Flink Optimizer: Logical Plan to
Execution Plan
If you want to know
one thing about Flink
is that you don't need
to know the internals of Flink
So … Batch
Batch on Stream
Stream: Unbounded Data Stream
Unbounded
Data Stream
Batch on Stream
Stream: Unbounded Data Stream
Batch: Bounded stream ( dataset ) on a stream processor
Global window over the entire dataset
Optimization in operators for joins and grouping,
with blocking data exchange if needed
Unbounded
Data Stream
Bounded
Data Set
Batch on Stream
Stream: Unbounded Data Stream
Batch: Bounded stream ( dataset ) on a stream processor
Global window over the entire dataset
Optimization in operators for joins and grouping,
with blocking data exchange if needed
Unbounded
Data Stream
Bounded
Data Set
Batch on Stream
Stream: Unbounded Data Stream
Batch: Bounded stream ( dataset ) on a stream processor
Global window over the entire dataset
Optimization in operators for joins and grouping,
with blocking data exchange if needed
Batch specific optimizations:
Cost-based optimizer: dataset size known before hand
Manage memory on / off-heap for join, sort, …
Optimization serialization stack for user-types
Bounded
Data Set
Batch on Stream
Unbounded
Data Stream
Conclusions
Conclusions
Conclusions
Flink Pure streaming engine matches real life. No Abstraction
Conclusions
Flink Pure streaming engine matches real life. No Abstraction
Batch on streaming
Conclusions
Flink Pure streaming engine matches real life. No Abstraction
Batch on streaming
Flexible Windowing Semantics with Explicit Time handling
Conclusions
Flink Pure streaming engine matches real life. No Abstraction
Batch on streaming
Flexible Windowing Semantics with Explicit Time handling
Competitive Performance, low latency and hight throughput
Conclusions
Flink Pure streaming engine matches real life. No Abstraction
Batch on streaming
Flexible Windowing Semantics with Explicit Time handling
Competitive Performance, low latency and hight throughput
Apache Beam, open sourced by Google, uses Flink as its first order runner for
Batch and Streaming processing in partnership with Data Artisans.
100% Compliance of data processing model “what, where, when, how “
¡gracias
!
136

Flink. Pure Streaming

  • 1.
    Flink Pure Streaming Paco Guerrero BigData & Solutions Architect 9/21/16
  • 2.
  • 4.
  • 5.
  • 6.
  • 7.
  • 8.
    “Abstraction of realityused to facilitate information processing” Batch
  • 9.
  • 10.
  • 11.
  • 12.
  • 13.
    Batch Batch Job All Input All Output Nothing abouttime Timestamps used as trick to keep real time fingerprint
  • 14.
    Streaming “Continuous processing ofdata that is continuously produced”
  • 15.
    Streaming “Streaming is thenext programming paradigm for data applications, and you need to start thinking in terms of streams” “Continuous processing of data that is continuously produced”
  • 16.
    Streaming “Streaming is thenext programming paradigm for data applications, and you need to start thinking in terms of streams” “Continuous processing of data that is continuously produced” Data Stream: Infinite sequence of data arriving in a continuous fashion.
  • 17.
    Streaming “Streaming is thenext programming paradigm for data applications, and you need to start thinking in terms of streams” “Continuous processing of data that is continuously produced” Data Stream: Infinite sequence of data arriving in a continuous fashion. Stream processing is the backbone of the new data infrastructure.
  • 18.
    Streaming “Streaming is thenext programming paradigm for data applications, and you need to start thinking in terms of streams” “Continuous processing of data that is continuously produced” Data Stream: Infinite sequence of data arriving in a continuous fashion. Stream processing is the backbone of the new data infrastructure. “The world beyond batch” A high-level tour of modern data-processing concepts. By Tyler Akidau August 5, 2015 https://www.oreilly.com/ideas/the-world-beyond-batch-streaming-101
  • 19.
  • 20.
  • 21.
  • 22.
  • 23.
    Streaming is thebiggest change in data infraestructure since Hadoop Streaming
  • 24.
    The biggest changeis moving from batch to streaming is handling time explicitly Streaming
  • 25.
  • 26.
  • 27.
  • 28.
    Batch Job Batch Job2 All Output Batch Job 1 Micro Batch
  • 29.
    Batch Job Batch Job2 All Input All OutputBatch Job 3 All Output All Output Batch Job 1 Batch Frequency ? Timestamps keeps real time fingerprint Micro Batch
  • 30.
    Streaming Technologies Batch StreamingMicroBatch StateLess – Record acknowledgements CPU bounded performance Not expressive declarative functional API – Low Level API Not auto scaling Low level programmatic topology Poor Streaming Windows funcionalities Not compatible with Hadoop APIs Streams
  • 31.
    Streaming Technologies Batch StreamingMicroBatch StateLess – Record acknowledgements CPU bounded performance Not expressive declarative functional API – Low Level API Not auto scaling Low level programmatic topology Poor Streaming Windows funcionalities Not compatible with Hadoop APIs Streams
  • 32.
    Streaming Technologies Batch StreamingMicroBatch StateLess – Record acknowledgements CPU bounded performance Not expressive declarative functional API – Low Level API Not auto scaling Low level programmatic topology Poor Streaming Windows funcionalities Not compatible with Hadoop APIs Streams
  • 33.
    Streaming Technologies Batch StreamingMicroBatch StateLess – Record acknowledgements CPU bounded performance Not expressive declarative functional API – Low Level API Not auto scaling Low level programmatic topology Poor Streaming Windows funcionalities Not compatible with Hadoop APIs Streams
  • 34.
  • 35.
  • 36.
  • 37.
  • 38.
  • 39.
    Flink Open Source StreamProcessing Framework. Last available Release 1.1.1 Top Level Apache Project since Dec '14
  • 40.
    Flink Open Source StreamProcessing Framework. Last available Release 1.1.1 Top Level Apache Project since Dec '14 Main Features Native Stream Low Latency High throughput Stateful Exactly-one guarantees Distributed Expressive APIs And more ….
  • 41.
    Flink Open Source StreamProcessing Framework. Last available Release 1.1.1 Top Level Apache Project since Dec '14 Main Features Native Stream Low Latency High throughput Stateful Exactly-one guarantees Distributed Expressive APIs And more ….
  • 42.
  • 43.
  • 44.
  • 45.
  • 46.
  • 47.
  • 48.
  • 49.
  • 50.
  • 51.
  • 52.
    Distributed pipelined processing Executeeverything as Stream Iterative ( cyclic ) dataflows Mutable state in operations Operate on managed memory (*) Also works on batch !! Job Manager Client Optimizer Dataflow Graph Flink Runtime Engine
  • 53.
    Distributed pipelined processing Executeeverything as Stream Iterative ( cyclic ) dataflows Mutable state in operations Operate on managed memory (*) Also works on batch !! Workers ( Task Managers ) Job Manager Client Optimizer Dataflow Graph Execution Graph Flink Runtime Engine
  • 54.
    Stream Job Batch Job MLJob Flink Runtime Engine Graph Job optimizer optimizer optimizer optimizer
  • 55.
    Stream Job Batch Job MLJob Flink Runtime Engine Graph Job optimizer optimizer optimizer optimizer
  • 56.
    Tasks scheduled andexecuted in workers ( slots ) Tasks as chain of operators Run operator logic in a pipelined fashion Stream Job Batch Job ML Job Flink Runtime Engine Graph Job optimizer optimizer optimizer optimizer
  • 57.
    Tasks scheduled andexecuted in workers ( slots ) Tasks as chain of operators Run operator logic in a pipelined fashion Stream Job Batch Job ML Job Flink Runtime Engine Graph Job optimizer optimizer optimizer optimizer
  • 58.
    Tasks scheduled andexecuted in workers ( slots ) Tasks as chain of operators Run operator logic in a pipelined fashion Stream Job Batch Job ML Job Flink Runtime Engine Graph Job optimizer optimizer optimizer optimizer
  • 59.
    If you wantto know one thing about Flink is that you don't need to know the internals of Flink
  • 60.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks
  • 61.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks lTime references lOut of order events lPowerful Windowing
  • 62.
    Event Times &Windowing
  • 63.
    Event Times &Windowing Event Time Event Time
  • 64.
    Event Times &Windowing Flink Data Source Event Time Event Time Ingestion Time
  • 65.
    Event Times &Windowing Flink Data Source Flink Window Operator Event Time Event Time Ingestion Time Processing Time
  • 66.
    Event Time: whendata is generated Ingestion time: when data is loaded from source Processing time: when data is processed Event time help to process out- of-order events and replay elements as the ocurred ( deterministic results ) Explicit handling of time. 3 choices: Event Times & Windowing
  • 67.
    Event Times &Windowing
  • 68.
  • 69.
    1 2 35 7 4 6 8 9 10 Event time. Out or Order
  • 70.
    1 2 35 7 4 6 8 9 10 Event time. Out or Order Out or Order 1 2 3 5 74 6 8 9 10
  • 71.
    1 2 35 7 4 6 8 9 10 1 2 3 5 74 6 8 9 104 Event time. Out or Order Ingestion Time WindowsOut or Order 1 2 3 5 74 6 8 9 10
  • 72.
    1 2 35 7 4 6 8 9 10 1 2 3 5 74 6 8 9 10 1 2 3 4 4 5 Event time. Out or Order 6 7 8 9 10 Event Time Windows Ingestion Time WindowsOut or Order 1 2 3 5 74 6 8 9 10
  • 73.
  • 74.
    1 2 35 7 4 6 8 9 10 Event time. Watermarks
  • 75.
    1 2 35 7 4 6 8 9 10 1 2 3 54 6 8 1 2 3 54 6 8 1 2 3 4 4 5 Event time. Watermarks 6 8 Event Time Windows Ingestion Time WindowsOut or Order
  • 76.
    1 2 35 7 4 6 8 9 10 1 2 3 54 6 8 910 1 2 3 54 6 8 910 1 2 3 4 4 5 Event time. Watermarks 6 8 9 10 Event Time Windows Ingestion Time WindowsOut or Order Not event time before 5 will come Late Time of 2 5
  • 77.
    1 2 35 7 4 6 8 9 10 1 2 3 5 74 6 8 910 1 2 3 5 74 6 8 910 1 2 3 4 4 5 Event time. Watermarks 6 7 8 9 10 Event Time Windows Ingestion Time WindowsOut or Order Not event time before 10 will come Late Time of 2 10
  • 78.
    Windowing Windows: grouping ofevents according to time, session*, count
  • 79.
    Windowing Windows: grouping ofevents according to time, session*, count Powerful built-in windows:
  • 80.
    Windowing Windows: grouping ofevents according to time, session*, count Powerful built-in windows: Count: number of events to trigger the window. Process X last events each Y events.
  • 81.
    Windowing Windows: grouping ofevents according to time, session*, count Powerful built-in windows: Count: number of events to trigger the window. Process X last events each Y events. Time: lTumbling: trigger every X time with received events lSliding: trigger every X time with received events in last Y time
  • 82.
    Windowing Windows: grouping ofevents according to time, session*, count Powerful built-in windows: Count: number of events to trigger the window. Process X last events each Y events. Time: lTumbling: trigger every X time with received events lSliding: trigger every X time with received events in last Y time Session: all events from session/user X until session time expired ( Gap )
  • 83.
    Windowing Windows: grouping ofevents according to time, session*, count Powerful built-in windows: Count: number of events to trigger the window. Process X last events each Y events. Time: lTumbling: trigger every X time with received events lSliding: trigger every X time with received events in last Y time Session: all events from session/user X until session time expired ( Gap ) High level API for user windows: Window Assigner, Trigger, Evictor
  • 84.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks lManaged operator state for backup/recovery lSavepoints
  • 85.
  • 86.
    Stateful Streaming Op Op State StatelessStream Processing Stateful Stream Processing lBuilt-in internal state in each operator for exactly-once semantics lUser state can be declared in each operator to be saved locally in memory ( API, key/value pars ) lSnapshots: periodically local states in memory are persisted in lightweight distributed snapshots. No global pause !! lCheckpoint as global consistent point-in-time snapshot build by set of distributed snapshots. lPluggable state backend for snapshots: JobManager, HDFS, RocksDB lSavepoints: user-triggered retained checkpoint
  • 87.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks lExactly-once semantics with managed operator state lDistributed Snapshotting Algorithm
  • 88.
    Periodically Chandy-Lamport Snapshots “The global-state-detectionalgorithm is to be superimposed on the underlying computation: It must run concurrently with, but no alter, this underlying computation” . Triggers snapshots asynchronously . Embedded snapshots algorithm in stream of data ( barriers ) . No global pause, lightweight impact in performance Handling Checkpoints
  • 89.
    Periodically Chandy-Lamport Snapshots “The global-state-detectionalgorithm is to be superimposed on the underlying computation: It must run concurrently with, but no alter, this underlying computation” . Triggers snapshots asynchronously . Embedded snapshots algorithm in stream of data ( barriers ) . No global pause, lightweight impact in performance Handling Checkpoints
  • 90.
    Periodically Chandy-Lamport Snapshots “The global-state-detectionalgorithm is to be superimposed on the underlying computation: It must run concurrently with, but no alter, this underlying computation” . Triggers snapshots asynchronously . Embedded snapshots algorithm in stream of data ( barriers ) . No global pause, lightweight impact in performance Handling Checkpoints
  • 91.
    snapshot Job Manager Periodically pushes barriers for newstate New state X+1 Ack for Snapshot state X from Task N Handling Checkpoints
  • 92.
  • 93.
  • 94.
    snapshot Job Manager Handling Checkpoints AllAcks received Register Checkpoint for restore in case of fail
  • 95.
    Streaming Fault Tolerance Incase of fail, last global checkpoint is recovered ( recovery from partial checkpoint / individual snapshots is coming ) Need of stateful source like kafka to ensure end-to-end exactly-once semantic in case of fail. Kafka sink doesn't guarantee end-to-end exactly-once ( multiple writes in topic ) ( at least-once ) Semantics in Flink: At Least Once: never loses events, events might be reprocessed Exactly once: neither reprocessed nor lost events. Exactly once by default, with low impact in performance
  • 96.
    If you wantto know one thing about Flink is that you don't need to know the internals of Flink
  • 97.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks lPipelined runtime lLatency vs throughput tunning
  • 98.
    Exactly-once semantic withlow impact in performance Controllable checkpointing overhead Higher throughput using processing time Performance improvements thanks to: . operator chaining during optimization phase . own optimized serialization stack with code generation Performance Tunning
  • 99.
    Benchmark for “StreamingComputation” published by Yahoo. Dec 18, 2015 https://yahooeng.tumblr.com/post/135321837876/benchmarking-streaming-computation-engines-at Production use-case lcounting ad impressions group by campaign laggregations over a 10 second window lsave current aggregate value to Redis every second Streaming Benchmark
  • 100.
    Throughput vs LatencyGraph Throughput ( 1000 events / sec ) 99 Percentile Latency ( ms ) Not Operator combinig in Storm, more complicate topology, more steps for events and more overhead
  • 101.
    Apache Storm WithoutTrident lAt least once / Double counting after fail / Lost state after Failures lCPU bounded Apache Spark lLatency increase with throughput Apache Flink lExactly once / No double counting / No state loss lLimited by bandwidth between Kafka and Flink cluster l(1 GigE). lkafka brokers within Kafka Cluster ( 10 GigE ) lAchieved 15 million messages /sec l( before 3 million m/sec) with exactly once semantic 10,000,000 20,000,000 1 GigE 10 GigE Performance Tunning
  • 102.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks lHigh Level API lWide range of basic and advanced operators lJava , Scala. Python soon !!
  • 103.
  • 104.
    API Working on datastreams ( bounded ? )
  • 105.
    API Working on datastreams ( bounded ? ) Stream Processing: Explicit Handling of Time
  • 106.
    API Working on datastreams ( bounded ? ) Stream Processing: Explicit Handling of Time Java & Scala. Python coming. Java: Bean type classes vs Tuples with position addresses. Scala: case classes.
  • 107.
    API Working on datastreams ( bounded ? ) Stream Processing: Explicit Handling of Time Java & Scala. Python coming. Java: Bean type classes vs Tuples with position addresses. Scala: case classes. Operators: Sources: kafka, FileSystem, Cassandra … Sinks: Kafka, HDFS, Cassandra …. Transformations: Basic: map, flatmap, filter, grouping, iterate, project, join, cross, … Streaming: Windowing + Aggregations, Temporal Binary Iterative Stream operators
  • 108.
    API Working on datastreams ( bounded ? ) Stream Processing: Explicit Handling of Time Java & Scala. Python coming. Java: Bean type classes vs Tuples with position addresses. Scala: case classes. Operators: Sources: kafka, FileSystem, Cassandra … Sinks: Kafka, HDFS, Cassandra …. Transformations: Basic: map, flatmap, filter, grouping, iterate, project, join, cross, … Streaming: Windowing + Aggregations, Temporal Binary Iterative Stream operators DataStream<?> DataSet<?> Core API 1 implementation*, 2 interfaces
  • 109.
    Source Map Reduce Fliter JoinSum Sink Map Source Operators
  • 110.
    Source Map Reduce Fliter JoinSum Sink Map Source Operators
  • 111.
    Source Map Reduce Fliter JoinSum Sink Source Filter Operators
  • 112.
    Source Map Reduce Fliter JoinSum Sink Source Filter Operators
  • 113.
    Source Map Reduce Fliter JoinSum Sink Source Reduce Operators
  • 114.
    Source Map Reduce Fliter JoinSum Sink Source Reduce Operators
  • 115.
    Source Map Reduce Fliter JoinSum Sink Source Join Operators
  • 116.
    Source Map Reduce Fliter JoinSum Sink Source Join Operators
  • 117.
    Source Map Reduce Fliter JoinSum Sink Source Operators
  • 118.
    Events Time & Windows Fault Tolerance & Correctness StateHandling Low Latency & High Throughput API Libraries SQL Building Blocks lEasy to use. SQL !! lBased on Apache Calcite
  • 119.
    API extension forDataSets y DataStreams Based on relational Table abstraction Table <=> Source / DataSet / DataStream Operators like: where, select, as, groupBy, join, union, minus, distinct, orderBy, ... Table API
  • 120.
    Execute SQL-Like sentenceson DataSets and Datastreams Resuts returned as Table ( Table API ), convertible to DataStream or DataSets SQL and Table API can be seamlessly mixed over DataStream/DataSets Flink’s SQL support is not feature complete, yet. Queries that include unsupported SQL will fail !! SQL Support SQL
  • 121.
    Parsing and Logicalplan for Table operators and SQL are optimized using Apache Calcite Only supported a Subset of the comprehensive SQL standard Apache Calcite provides with: SQL Parsing API for building expressions in relational algebra Query planning engine Provides SQL for Streaming Queries with windows aggregations SELECT STREAM TUMBLE_END(rowtime, INTERVAL '1' HOUR) AS rowtime, productId, COUNT(*) AS c, SUM(units) AS units FROM Orders Apache Calcite SQL Sentence Apache Calcite: SQL to Logical Plan as Relational Algebra Flink Optimizer: Logical Plan to Execution Plan
  • 122.
    If you wantto know one thing about Flink is that you don't need to know the internals of Flink So … Batch
  • 123.
  • 124.
    Stream: Unbounded DataStream Unbounded Data Stream Batch on Stream
  • 125.
    Stream: Unbounded DataStream Batch: Bounded stream ( dataset ) on a stream processor Global window over the entire dataset Optimization in operators for joins and grouping, with blocking data exchange if needed Unbounded Data Stream Bounded Data Set Batch on Stream
  • 126.
    Stream: Unbounded DataStream Batch: Bounded stream ( dataset ) on a stream processor Global window over the entire dataset Optimization in operators for joins and grouping, with blocking data exchange if needed Unbounded Data Stream Bounded Data Set Batch on Stream
  • 127.
    Stream: Unbounded DataStream Batch: Bounded stream ( dataset ) on a stream processor Global window over the entire dataset Optimization in operators for joins and grouping, with blocking data exchange if needed Batch specific optimizations: Cost-based optimizer: dataset size known before hand Manage memory on / off-heap for join, sort, … Optimization serialization stack for user-types Bounded Data Set Batch on Stream Unbounded Data Stream
  • 128.
  • 129.
  • 130.
    Conclusions Flink Pure streamingengine matches real life. No Abstraction
  • 131.
    Conclusions Flink Pure streamingengine matches real life. No Abstraction Batch on streaming
  • 132.
    Conclusions Flink Pure streamingengine matches real life. No Abstraction Batch on streaming Flexible Windowing Semantics with Explicit Time handling
  • 133.
    Conclusions Flink Pure streamingengine matches real life. No Abstraction Batch on streaming Flexible Windowing Semantics with Explicit Time handling Competitive Performance, low latency and hight throughput
  • 134.
    Conclusions Flink Pure streamingengine matches real life. No Abstraction Batch on streaming Flexible Windowing Semantics with Explicit Time handling Competitive Performance, low latency and hight throughput Apache Beam, open sourced by Google, uses Flink as its first order runner for Batch and Streaming processing in partnership with Data Artisans. 100% Compliance of data processing model “what, where, when, how “
  • 136.